My only question is whether the original optimization is improving real-world performance enough to justify the considerable complexity it introduces.

I haven't looked to see exactly which optimization this is, but IIRC, it
provides a significant performance difference between scatter() and
plot()-without-lines. Up until two(?) years ago, the going recommendation
for plotting many scatter points was to do plot() with a blank linestyle
because its marker optimization was not available to scatter() (because
scatter can have varying markers). Multiple projects approached me during
SciPy 2011 about the performance issue (particularly networkx) and
switching to plot() solved the issue for them. mplot3d also used to show
significant performance issues when doing scatter() (don't know if it still
does) because redrawing interactively took longer.

On Fri, Nov 21, 2014 at 12:38 PM, Eric Firing notifications@github.com
wrote:

My only question is whether the original optimization is improving
real-world performance enough to justify the considerable complexity it
introduces.

—
Reply to this email directly or view it on GitHub
#3826 (comment)
.

I think this was originally implemented at the request of @ChrisBeaumont of the Glue project. It uses scatter extensively, and was unusably slow without this optimization.

OK, that is awakening some traces in my memory. It makes sense. While you are in the code, @mdboom, would you consider adding a comment, please? Something like "Optimization: use fast marker rendering when possible. This is needed for some uses of scatter, for example."

👍 LGTM